fix(cuda): restore DeepSeek-V2 output by reverting the regressed RMSNorm kernel

.github/workflows/release.yml

@@ -29,7 +29,7 @@ permissions: {}
 # documented in CLAUDE.md). Hoisted to workflow scope so a future bump
 # touches one literal here instead of one per job.
 env:
-  MLX_EXPECTED_COMMIT: "57c66cac7cb3e5b1eb350488a61f1506b40d39f8"
+  MLX_EXPECTED_COMMIT: "b7c3dd6d27f45b5365b08a840310187dc503f1db"
 
 jobs:
   # ============================================================================

src/distributed/pipeline/stage_executor/mod.rs

@@ -168,15 +168,16 @@ impl LoadedStageExecutor {
             stage.stage_index
         );
 
-        // Issue #688 (M2 hardening): defense-in-depth CUDA-graph disable for a
-        // Gemma 4 pipeline stage, mirroring the non-PP backend-seam load sites.
+        // Issue #688 (M2 hardening), extended to DeepSeek-V2 by issue #824:
+        // defense-in-depth CUDA-graph disable for a hazard-family pipeline stage,
+        // mirroring the non-PP backend-seam load sites.
         // Both the in-process and the remote-process pipeline loaders reach this
         // common stage-load chokepoint before any stage weights are realised. The
         // authoritative env write already happens on the main startup thread in
         // `start_server`; this idempotent call (guarded by `var_os`) only takes
         // effect if a future pipeline entry ever bypasses that hoist, so the env is
         // set before the stage's first GPU eval regardless.
-        crate::loading::maybe_disable_cuda_graphs_for_gemma4_for_path(model_dir);
+        crate::loading::maybe_disable_cuda_graphs_for_model_for_path(model_dir);
 
         let filter = LayerFilter::from_stage(stage);
         let family = resolve_stage_family(model_dir)?;

src/lib/mlx-cpp/CMakeLists.txt

@@ -92,7 +92,7 @@ else()
   FetchContent_Declare(
     mlx
     GIT_REPOSITORY "https://github.com/ml-explore/mlx.git"
-    GIT_TAG 57c66cac7cb3e5b1eb350488a61f1506b40d39f8)
+    GIT_TAG b7c3dd6d27f45b5365b08a840310187dc503f1db)
 
   # Use FetchContent_Populate + add_subdirectory so we can apply source
   # overlays before the MLX build system processes the files.